PeopleSearchBench results
We ran Happenstance on all 119 queries of PeopleSearchBench, the open benchmark for AI people search, and scored 84.7 out of 100.
By the Happenstance team · Ran October 27, 2026
The benchmark
PeopleSearchBench was published this year by LessieAI. Its paper and code are public. It has 119 queries across four use cases, which the benchmark calls scenarios, in English, Portuguese, Spanish and Dutch.
It turns each query into a checklist of criteria, and an AI judge uses web search to check the people a platform returns against that list. The overall score is the mean of three measures, each out of 100:
- Relevance: were the right people at the top of the list?
- Coverage: what share of the top 10 were qualified people?
- Utility: how complete and useful was each profile?
| Overall | Relevance | Coverage | Utility | |
|---|---|---|---|---|
| Happenstance | 84.7 | 81.9 | 89.3 | 82.8 |
| Lessie | 65.2 | 70.2 | 69.1 | 56.4 |
| Exa | 55.0 | 53.8 | 58.1 | 53.1 |
| Claude Code | 46.0 | 54.3 | 41.1 | 42.7 |
| Juicebox | 45.8 | 44.7 | 41.8 | 50.9 |
We led on all three measures. The widest gap was utility, at 26.4 points. Our profiles were more complete and gave more evidence for why each person matched.
Results by use case
The benchmark covers four use cases:
- Recruiting: candidates with specific skills, experience and location.
- B2B prospecting: customers and decision makers at certain kinds of companies.
- Expert search: the leading people in a field, such as researchers in AI safety.
- Influencers: creators in a niche with a set audience size, such as 10K to 80K followers.
Happenstance was ahead of every platform on the leaderboard in all four.
Our lead is widest in B2B prospecting, at 29.4 points, and narrowest in influencers, at 8.7. Influencers is also our lowest score. Most of those queries ask for a follower count, which our search doesn't track yet.
How we ran it
- Our search
- Our production search over our full index of people, one run per query, on October 27, 2026. We did zero tweaks to improve against this benchmark (i.e. no “benchmaxxing”). We report our system's performance as is. We disabled our input screener, so it ran even on test cases our system is ill-equipped to deal with, such as filtering influencers by number of followers.
- The judge
- Gemini 3 Flash Preview, using the benchmark's public code (commit f1afbb9) and the paper's settings, with its prompts unchanged. We called Gemini through Google's own client, because the code's OpenRouter client fails on Gemini 3 tool calls.
- What it saw
- Our top 15 people per query, as the submission guide recommends, with the top 10 scored. For each person: the profile and contact details we hold, our summary of why they match, and the evidence for each criterion.
- Other platforms
- Scores as LessieAI published them. We didn't rerun them, so they were judged earlier than our run.
- Per-person results
- We aren't publishing them, because they name real people. The benchmark's own repo leaves them out for the same reason.
Find who you can reach
PeopleSearchBench asks for anyone in the world who fits a query, so we searched our full index. On Happenstance, the same search engine runs over your network: your connections, your friends' connections, and your groups. Every result comes with how you're connected.
The benchmark tests how well we find anyone in the world. Happenstance is built to find the people you can reach. Join 500,000+ professionals already searching their whole network.
Continue with Google